OrcaRouter

OpenAI-compatible LLM routing gateway with self-hosted Lite edition and hosted routing features for model selection, fan-out, arbiters, DSL rules, shadowing, and canary rollout.

What It Is

OrcaRouter is not OpenRouter. Its useful signal for Kevin is the programmable routing layer: keep one application-facing model slug, then change the graph behind it. The launch thread frames Fusion as a model panel where a prompt fans out to several models, a judge reads the answers, and one strongest answer returns through a single API. Source: X/@OrcaRouter, 2026-06-15

The self-hosted Lite repository is a single-workspace, BYOK, OpenAI-compatible router with a managed fallback. GitHub describes it as Python, MIT licensed, created 2026-05-03, last pushed 2026-05-28, with 510 stars and 49 forks as of 2026-06-30. Source: GitHub API, 2026-06-30

Routing DSL

The stronger architectural idea is the hosted Routing DSL: YAML rules with CEL conditions evaluate request shape, task classification, headers, tenant context, and agent session state. Effects can route to a model, model list, pool, or named router; parallel effects can fan out and pass answers to an arbiter. Source: OrcaRouter docs, 2026-06-30

Important primitives:

• built-in strategies for cheapest, quality, balanced, and adaptive routing;
• DSL rules for task specialization, difficulty-aware escalation, agent-aware repair, and tenant/time/header routing;
• agent_state variables such as turn count, tools used, files read, whether edits happened, whether tests failed, and consecutive errors;
• arbiters including best_of_n, synthesize, majority, first, and tests_pass;
• safety controls: linting, dry-runs, version history, audit rows, shadow mode, canary percentages, and rollback.

This is LLM Council turned into infrastructure. It gives the operator a programmable boundary between app code and model choice, and lets the routing graph evolve without redeploying the app.

Stack Fit

Use OrcaRouter as an idea source for Dedalus-style model routing and eval-gated fan-out. Do not adopt it as a default Kevin dependency without checking data retention, provider key custody, hosted-vs-self-hosted boundaries, and whether the DSL's agent-state variables map to the local harness.

For local brain work, the durable thought is: "frontier model" can be a graph the operator composes. The real asset is the versioned routing policy plus evaluation receipts, not the marketing slug.

---

Timeline

• 2026-06-30 | Page created from OrcaRouter Fusion launch bookmark, Routing DSL docs, and OrcaRouter Lite repository metadata. Source: X/@OrcaRouter, 2026-06-15; OrcaRouter docs, 2026-06-30
• 2026-06-15 | Launch bookmark framed OrcaRouter Fusion as a one-call model panel with a judge and a DSL for custom routing graphs. Source: X/@OrcaRouter, 2026-06-15